2nd/10_Wiki/Topics/DevOps_and_Security/Frustum Culling.md

---
id: wiki-2026-0508-frustum-culling
title: Frustum Culling
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [View Frustum Culling, VFC, Camera Culling]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [graphics, rendering, culling, optimization, gpu]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: cpp
  framework: opengl-vulkan-unity
---

# Frustum Culling

## 매 한 줄
> **"매 carmera 의 view volume (frustum) 밖 object 의 매 draw skip"**. 매 가장 기본적이고 가장 효과적인 매 visibility culling — 매 30-90% draw call 감소가 일반적. 매 modern engine (Unreal 5 Nanite, Unity HDRP, bevy) 은 매 GPU-driven culling 으로 매 millions of objects 를 매 compute shader 안에서 매 frame 마다 cull.

## 매 핵심

### 매 frustum 표현
- **6 planes**: near, far, left, right, top, bottom.
- 매 plane equation: `ax + by + cz + d = 0` with `(a,b,c)` = inward normal.
- 매 view-projection matrix 의 매 row combo 로 6 planes extract (Gribb-Hartmann).

### 매 bounding volume choice
- **AABB (axis-aligned)**: 매 cheapest, 매 conservative — 매 large rotated objects 매 over-conservative.
- **OBB (oriented)**: 매 tighter, 매 더 expensive.
- **Sphere**: 매 cheapest test (single dot product), 매 loosest.
- **Plane mask (frustum culling with masks)**: 매 children inherit parent 의 "fully inside" plane.

### 매 알고리즘 흐름
1. View-projection matrix → 6 frustum planes.
2. 매 object 의 BV 와 매 6 planes test.
3. **Outside** any plane → cull.
4. **Inside** all → render.
5. **Intersect** → render (or recurse children if hierarchy).

### 매 modern (GPU-driven)
- **Compute shader** 가 매 draw arguments buffer 를 build (`DrawIndirect`).
- 매 millions of objects 도 매 sub-millisecond.
- 매 hierarchical Z-buffer occlusion + frustum 결합 (Nanite).

## 💻 패턴

### Extract frustum planes from VP matrix (Gribb-Hartmann)
```cpp
struct Plane { glm::vec3 n; float d; };

void extractPlanes(const glm::mat4& vp, Plane out[6]) {
    auto m = glm::transpose(vp); // row-major helper
    out[0] = { glm::vec3(m[3]+m[0]), m[3].w + m[0].w }; // left
    out[1] = { glm::vec3(m[3]-m[0]), m[3].w - m[0].w }; // right
    out[2] = { glm::vec3(m[3]+m[1]), m[3].w + m[1].w }; // bottom
    out[3] = { glm::vec3(m[3]-m[1]), m[3].w - m[1].w }; // top
    out[4] = { glm::vec3(m[3]+m[2]), m[3].w + m[2].w }; // near
    out[5] = { glm::vec3(m[3]-m[2]), m[3].w - m[2].w }; // far
    for (int i = 0; i < 6; i++) {
        float len = glm::length(out[i].n);
        out[i].n /= len; out[i].d /= len;
    }
}
```

### Sphere vs frustum (cheapest)
```cpp
bool sphereInFrustum(const Plane planes[6], const glm::vec3& c, float r) {
    for (int i = 0; i < 6; i++)
        if (glm::dot(planes[i].n, c) + planes[i].d < -r) return false;
    return true;
}
```

### AABB vs frustum (positive vertex / p-vertex test)
```cpp
bool aabbInFrustum(const Plane planes[6], const glm::vec3& mn, const glm::vec3& mx) {
    for (int i = 0; i < 6; i++) {
        glm::vec3 p = {
            planes[i].n.x >= 0 ? mx.x : mn.x,
            planes[i].n.y >= 0 ? mx.y : mn.y,
            planes[i].n.z >= 0 ? mx.z : mn.z
        };
        if (glm::dot(planes[i].n, p) + planes[i].d < 0) return false;
    }
    return true;
}
```

### BVH-based hierarchical culling
```cpp
void cullBVH(const BVHNode& node, const Plane planes[6], std::vector<int>& visible) {
    auto r = aabbVsFrustumIntersect(planes, node.aabb);
    if (r == OUTSIDE) return;
    if (r == INSIDE)  { addAll(node, visible); return; }
    if (node.isLeaf) {
        for (int idx : node.objects)
            if (aabbInFrustum(planes, objs[idx].mn, objs[idx].mx))
                visible.push_back(idx);
        return;
    }
    cullBVH(*node.left,  planes, visible);
    cullBVH(*node.right, planes, visible);
}
```

### GPU compute culling (HLSL)
```hlsl
// CullCS.hlsl
StructuredBuffer<ObjectData> objects : register(t0);
ConstantBuffer<FrustumCB>    frustum : register(b0);
RWStructuredBuffer<DrawArgs> drawArgs : register(u0);
RWByteAddressBuffer          counter  : register(u1);

[numthreads(64, 1, 1)]
void main(uint3 id : SV_DispatchThreadID) {
    if (id.x >= objects.Length) return;
    ObjectData o = objects[id.x];
    bool visible = true;
    [unroll] for (int i = 0; i < 6; i++) {
        float4 p = frustum.planes[i];
        if (dot(p.xyz, o.center) + p.w < -o.radius) { visible = false; break; }
    }
    if (visible) {
        uint slot;
        counter.InterlockedAdd(0, 1, slot);
        drawArgs[slot].vertexCount    = o.indexCount;
        drawArgs[slot].instanceCount  = 1;
        drawArgs[slot].firstIndex     = o.firstIndex;
        drawArgs[slot].baseInstance   = id.x;
    }
}
```

### Unity (Burst) culling job
```csharp
[BurstCompile]
struct FrustumCullJob : IJobParallelFor {
    [ReadOnly] public NativeArray<float4> planes;       // 6 planes
    [ReadOnly] public NativeArray<float4> bounds;       // xyz=center, w=radius
    [WriteOnly] public NativeArray<bool>  visible;

    public void Execute(int i) {
        float4 b = bounds[i];
        for (int p = 0; p < 6; p++) {
            float4 pl = planes[p];
            if (math.dot(pl.xyz, b.xyz) + pl.w < -b.w) {
                visible[i] = false;
                return;
            }
        }
        visible[i] = true;
    }
}
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| <1k objects, CPU | per-object sphere/AABB test |
| 1k-100k, hierarchical | BVH / Octree + frustum |
| 100k+ static, GPU | compute shader + DrawIndirect |
| Massive (Nanite-class) | GPU-driven + HZB occlusion |
| Animated skeletal | use skinned bounds (loose) |

**기본값**: 매 modern engine — GPU compute culling + BVH for spatial queries.

## 🔗 Graph
- 부모: [[Real-Time Rendering]]
- 응용: [[GPU-Driven Rendering]] · [[Nanite]]
- Adjacent: [[BVH]] · [[Octree]]

## 🤖 LLM 활용
**언제**: plane extraction code 검토, false-cull bug 디버깅 (e.g., flipped normal), GPU shader skeleton.
**언제 X**: 매 actual rendering decision 의 runtime correctness — unit test + visual verification.

## ❌ 안티패턴
- **No bounding volume cache**: 매 frame 마다 매 mesh 의 bound 재계산 — pre-compute.
- **Sphere only for everything**: 매 long thin object 매 over-conservative.
- **Plane normalization 누락**: 매 distance comparison 부정확.
- **Cull camera == render camera 가정**: 매 shadow camera, planar reflection 시 매 잘못.
- **Animated bound 무시**: 매 skinned mesh 의 bound 가 매 outdated → pop in/out.

## 🧪 검증 / 중복
- Verified (Real-Time Rendering 4th ed, Gribb-Hartmann 2001, Unreal Nanite docs 2026).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — frustum extraction + BV tests + GPU-driven |